[rfc][patch 1/2] inotify: fix race

[Nick Piggin <npiggin@suse.de>]

Hmm, thinking I'd better just take a hammer to this inotify problem
despite the fact I cannot confirm the root cause of the warning messages.
Possibilities include the race being fixed here, and races in the
debug code itself.

--

There is a race between setting an inode's children's "parent watched" flag
when placing the first watch on a parent, and instantiating new children of
that parent: a child could miss having its flags set by set_dentry_child_flags,
but then inotify_d_instantiate might still see !inotify_inode_watched.

The solution is to set_dentry_child_flags after adding the watch. Locking is
taken care of, because both set_dentry_child_flags and inotify_d_instantiate
hold dcache_lock and child->d_locks.

Signed-off-by: Nick Piggin <npiggin@suse.de>
---
Index: linux-2.6/fs/inotify.c
===================================================================
--- linux-2.6.orig/fs/inotify.c
+++ linux-2.6/fs/inotify.c
@@ -627,6 +627,7 @@ s32 inotify_add_watch(struct inotify_han
 		      struct inode *inode, u32 mask)
 {
 	int ret = 0;
+	int newly_watched;
 
 	/* don't allow invalid bits: we don't want flags set */
 	mask &= IN_ALL_EVENTS | IN_ONESHOT;
@@ -653,12 +654,18 @@ s32 inotify_add_watch(struct inotify_han
 	 */
 	watch->inode = igrab(inode);
 
-	if (!inotify_inode_watched(inode))
-		set_dentry_child_flags(inode, 1);
-
 	/* Add the watch to the handle's and the inode's list */
+	newly_watched = !inotify_inode_watched(inode);
 	list_add(&watch->h_list, &ih->watches);
 	list_add(&watch->i_list, &inode->inotify_watches);
+	/*
+	 * Set child flags _after_ adding the watch, so there is no race
+	 * windows where newly instantiated children could miss their parent's
+	 * watched flag.
+	 */
+	if (newly_watched)
+		set_dentry_child_flags(inode, 1);
+
 out:
 	mutex_unlock(&ih->mutex);
 	mutex_unlock(&inode->inotify_mutex);